Abstract: It is imperative that Alzheimer’s Disease is caught in its early stages to prevent rapid progression, but it is often difficult to be diagnosed both quickly and inexpensively. Given that speech degradation is one of the earliest symptoms of AD, it has been suggested that neural nets can be used to classify speech data for AD. In this study, a network with a bi-directional GRU and four dense layers was trained on a relatively limited dataset from DementiaBank with 243 samples in each category. The model was found to have a mean and maximum accuracy of 0.63 and 0.825 when randomly tested 40 times, and an AUC ROC score of 0.654 when cross-validated. While these values are not ideal, they prove that using RNNs for AD diagnosing is promising.